StealthMark: Harmless and Stealthy Ownership Verification for Medical Segmentation via Uncertainty-Guided Backdoors

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Vision

Background:

• Medical data annotation for AI training is expensive and limited by specialist shortages and privacy concerns.
• Well-trained medical segmentation models are valuable intellectual property but lack robust protection methods.
• Existing model protection techniques largely overlook segmentation models, crucial for medical image analysis.

Purpose of the Study:

• To propose StealthMark, a novel, stealthy, and harmless method for verifying ownership of medical segmentation models.
• To enable black-box ownership verification without compromising model performance.
• To address the underexplored area of protecting medical segmentation models.

Main Methods:

• StealthMark subtly modulates model uncertainty without altering segmentation outputs.
• Model-agnostic explanation methods (e.g., LIME) extract feature attributions for watermark revelation.
• A QR code is designed as a watermark for robust and recognizable ownership claims.

Main Results:

• StealthMark effectively verifies ownership of medical segmentation models across diverse datasets and models.
• The method is stealthy and harmless, maintaining original model performance (e.g., <1% drop in Dice/AUC for SAM model).
• Achieved high attack success rates (>95%) while preserving segmentation accuracy, outperforming backdoor methods.

Conclusions:

• StealthMark provides an effective, stealthy, and harmless solution for medical segmentation model ownership verification.
• The QR code watermark facilitates robust and recognizable ownership claims.
• The method shows strong potential for practical deployment in protecting valuable medical AI intellectual property.